Bellkor breaks 9% improvement in the Netflix Challenge (as of 4/1) | Hacker News Reader

Bellkor breaks 9% improvement in the Netflix Challenge (as of 4/1)

netflixprize.com

34 pointsdkasper12 comments

What has actually been accomplished? Details please.

(edited, brevity.)

The biggest accomplishment is the algorithm - when applied to other useful noisy datasets, it can make all the difference between winning and losing (e.g. financial markets)

http://www.research.att.com/~volinsky/netflix/ProgressPrize2...

That paper describes what they are doing.

The Wikipedia page is pretty decent:

http://en.wikipedia.org/wiki/Netflix_Prize

The class of algorithms that it's dealing with is called collaborative filtering and the Bell Labs team has done some impressive stuff there. The contest is more academic than it would seem on the surface, since as was mentioned in a recently posted article, the data set is rather narrow in scope. In other words, it's more of a collaborative filtering algorithms contest than a "find the best recommendation algorithm" contest.

Any chance to see it used here? Another menu at the top: recommended?

The Netflix prize algorithms rely heavily on large amounts of data (thousands of votes per movie). It might be useful for Reddit and Digg as each story gets votes in the 3-5 digits, but not so much for this site. Of course, it would definitely be something, but I don't think it would be as useful.

Not to mention, the Netflix data has a scale of 1-5. Digg-like sites have scales from 0-1 (no vote, or vote), so that might further dilute the quality.

Popular stuff is easy to predict, it is the users with little feedback and less popular movies that make all the difference. Digg doesn't need a recommender because it is all about popularity, and if you do not like it - do not go there. I also want to go back to see the best stories on news.YC before I joined and I do not see a way to do that.

Yeah, I've often thought that the ability to see historical YC news stories would be very useful -- both for the period before a user joined the site, as well as when you miss a week or two because of vacation, illness, etc.

For that you can use http://www.aiderss.com/best/news.ycombinator.com , it is filtered based on popularity

I have written a social news site which you might be interested in: http://news.ycombinator.com/item?id=135495

Congratulations, I have written one too, what a coincidence (sarcasm directed towards situation, not the person). I bet most of the people reading this thread have done something around netflix results :)

Having a "recommended" link is a hard problem for news sites like this one. If you have a well defined category then there are lots of techniques that will give you good results.

At a place like this, people are looking for "interesting tech articles", which is a really vague thing. Also, interesting articles tend to be things that you haven't seen before. So it is hard to build up a training set for that.

Reply on news.ycombinator.com